GPT-5.4 Mini Hits ChatGPT Free/Go — Efficiency Tier, Lab Benches
OpenAI put GPT-5.4 mini in ChatGPT Thinking for Free and Go users (March 18), a day after API/Codex mini+nano. API: $0.75/$4.50 per 1M in/out, 400k context. Lab benches (high/xhigh): SWE-Bench Pro 54.4%, OSWorld-Verified 72.1%, GPQA Diamond 88.0% — vendor numbers; also rate-limit fallback for higher tiers.
Flagship releases get the headlines; mini is how OpenAI fills Free/Go and rate-limit gaps. The trade: most of the 5.4 coding/reasoning story at lower latency and cost — lab benches still OpenAI’s.
OpenAI made GPT-5.4 mini available in ChatGPT (March 18) to Free and Go users via the Thinking option (+ menu). API and Codex mini/nano landed March 17; the GPT-5.4 family (Thinking/Pro) launched March 5. Mini targets responsive coding assistants, subagents, computer-use agents, and multimodal apps; also acts as rate-limit fallback for GPT-5.4 Thinking on higher tiers. Nano remains API-only at lower pricing.
| Claim (OpenAI, high/xhigh) | Figure |
|---|---|
| SWE-Bench Pro (Public) | 54.4% |
| OSWorld-Verified | 72.1% |
| GPQA Diamond | 88.0% |
| API price | $0.75 / $4.50 per 1M in/out |
| Context | 400k |
| I/O | Text/image; tools, function calling, web/file search, computer use, skills |
Claims vs checks
Availability and pricing are OpenAI primary (blog March 17; Help Center release notes March 18). Benchmarks are vendor at stated effort levels — independent Artificial Analysis / Arena Elo not required for this ChatGPT-surface story but should be checked before ranking freezes.
Limits
- Free/Go Thinking quotas still apply.
- Effort-skewed benches are not default-chat Elo.
- Nano economics differ; do not conflate mini with nano.
Sources
- OpenAI: “Introducing GPT-5.4 mini and nano” (March 17, 2026).
- OpenAI Help: Model Release Notes — GPT-5.4 mini in ChatGPT (March 18, 2026).
- Community and benchmark tracking confirming rollout timing.